We are going to pay through the nose to get access to the information we all collectively gathered and created on sites like WikiPedia, Reddit, StackOverflow, and more.
All that free labour we all did to write and curate this information... scraped and locked behind the coming #AI pricing explosion.
@syntaxseed Yeah. I do hope people will contribute to open source AI tho
Jul 3, 2026, 7:23 PM@syntaxseed So one interesting thing here is the subsidies are relative to "API pricing", not necessarily OpenAI/Anthropic's marginal cost per token.
My guess is that OpenAI doesn't make much on someone using a $200 Codex plan flat-out, but that's after a ~40x subsidy vs. pay-per-token.
We don't know how big their models are, but serving costs at scale for bigger open-weights models are on the order of 99% less than GPT-5.5 costs.
Inference being a money-printer is why Nvidia can charge $$$.
Jul 3, 2026, 7:52 PM@syntaxseed my back-of-envelope first impression after being involved in a model-training effort is "it takes as much labour to train an LLM to do a task as it does to write a deterministic algorithm for the same task". The only reason executives are able to delude themselves that LLM tech is financially viable is because the training data is usually stolen from the community that produced it.
@syntaxseed this site is hacked and tried to get me to download malware.